Add files using upload-large-folder tool

.gitattributes

@@ -35,3 +35,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
 model.onnx.data filter=lfs diff=lfs merge=lfs -text
+onnx/model_q4f16.onnx_data_1 filter=lfs diff=lfs merge=lfs -text
+onnx/model_q4f16.onnx_data filter=lfs diff=lfs merge=lfs -text
+onnx/model_q4f16.onnx_data_2 filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,22 +1,25 @@
 ---
 license: mit
 base_model: inclusionAI/Ling-3.0-tiny
-library_name: onnxruntime
+library_name: transformers.js
 pipeline_tag: text-generation
 tags:
   - onnx
+  - transformers.js
+  - webgpu
   - onnxruntime
   - onnxruntime-genai
   - q4
+  - q4f16
   - int4
   - cuda
   - moe
   - text-generation
 ---
 
-# Ling-3.0-tiny ONNX — Q4 CUDA
+# Ling-3.0-tiny ONNX — Q4 WebGPU + CUDA
 
-CUDA-targeted ONNX Runtime GenAI conversion of
+Community ONNX conversion of
 [`inclusionAI/Ling-3.0-tiny`](https://huggingface.co/inclusionAI/Ling-3.0-tiny),
 a 7.9B-total / 1.3B-active-parameter hybrid reasoning MoE model.
 
@@ -25,52 +28,97 @@ a 7.9B-total / 1.3B-active-parameter hybrid reasoning MoE model.
 > release. Read the [original model card](https://huggingface.co/inclusionAI/Ling-3.0-tiny)
 > for training, evaluation, intended use, and base-model limitations.
 
-| Item | Value |
-|---|---|
-| Base model | [`inclusionAI/Ling-3.0-tiny`](https://huggingface.co/inclusionAI/Ling-3.0-tiny) |
-| Runtime | ONNX Runtime GenAI with CUDA |
-| Quantization | Q4 weight-only, block size 32 |
-| Activations / cache | FP16 |
-| Router math | FP32 |
-| Context length | 131,072 tokens, subject to available VRAM |
-| Download size | 4.849 GB / 4.516 GiB |
+| Target | Files | Runtime | Download |
+|---|---|---|---:|
+| Browser WebGPU | `onnx/model_q4f16.onnx` + 3 data shards | Transformers.js 4.2+ | 4.849 GB / 4.516 GiB |
+| NVIDIA CUDA | `model.onnx` + `model.onnx.data` | ONNX Runtime GenAI | 4.849 GB / 4.516 GiB |
 
-## Architecture preserved
+The WebGPU layout follows the standard Transformers.js `q4f16` contract used
+by browser-oriented ONNX repositories: the graph is under `onnx/`, external
+tensor data is split into three sub-2 GB files, activations and cache are FP16,
+and hybrid recurrent-cache names use the Qwen3.5-compatible convention already
+supported by Transformers.js.
 
-The conversion keeps Ling's hybrid architecture rather than replacing its
-specialized layers with dense approximations:
+## Run in the browser with WebGPU
 
-- 18 Kimi Delta Attention (KDA) layers with recurrent and convolution state
-- 6 Multi-Latent Attention (MLA) layers with KV cache
-- 23 sparse MoE layers with 128 routed experts, top-8 group-limited routing,
-  expert bias, and one shared expert
-- one dense MLP layer at the start of the decoder
+Install Transformers.js 4.2 or newer:
 
-The graph uses CUDA `MatMulNBits`, `QMoE`, `LinearAttention`,
-`CausalConvWithState`, and `GroupQueryAttention` operators. `model.onnx` and
-`model.onnx.data` form one model and must remain in the same directory.
+```bash
+npm install "@huggingface/transformers@^4.2.0"
+```
 
-## Install
+```javascript
+import { pipeline, TextStreamer } from '@huggingface/transformers';
+
+const generator = await pipeline(
+  'text-generation',
+  'webbrain-one/Ling-3.0-tiny-ONNX',
+  {
+    device: 'webgpu',
+    dtype: 'q4f16',
+  },
+);
+
+const messages = [
+  { role: 'user', content: 'Explain why the sky is blue in two sentences.' },
+];
+
+const output = await generator(messages, {
+  max_new_tokens: 128,
+  do_sample: false,
+  tokenizer_encode_kwargs: { enable_thinking: false },
+  streamer: new TextStreamer(generator.tokenizer, {
+    skip_prompt: true,
+    skip_special_tokens: true,
+  }),
+});
+
+console.log(output[0].generated_text.at(-1)?.content);
+```
 
-Python 3.10 or newer is recommended. This graph requires ONNX Runtime 1.28.0+
-and ONNX Runtime GenAI 0.15.2+.
+For thinking mode, set `enable_thinking: true`. The original model card
+recommends `temperature: 1.0`, `top_p: 0.95`, and `top_k: 20` when sampling in
+thinking mode.
 
-```bash
-pip install "onnxruntime-gpu>=1.28.0" \
-  "onnxruntime-genai-cuda>=0.15.2" \
-  "transformers>=4.57,<5"
+### Browser requirements
+
+- A current desktop browser with WebGPU enabled; Chrome or Edge is recommended.
+- Approximately 4.85 GB of model downloads on first load, plus browser cache.
+- Enough GPU memory for all weights, runtime buffers, state, and the requested
+  context. The 1.3B active-parameter figure reduces compute, but all 7.9B model
+  parameters still need to be stored.
+- Start with a short prompt and modest `max_new_tokens`, then increase context
+  after confirming memory use on the target device.
+
+This graph depends on WebGPU implementations of `MatMulNBits`, `QMoE`,
+`LinearAttention`, `CausalConvWithState`, and `GroupQueryAttention`. It is not a
+WASM/CPU fallback model.
+
+## Use from WebBrain
+
+Use the same repository ID and standard Transformers.js settings:
+
+```text
+model: webbrain-one/Ling-3.0-tiny-ONNX
+device: webgpu
+dtype: q4f16
+task: text-generation
 ```
 
-Download all repository files into one directory:
+WebBrain should select `onnx/model_q4f16.onnx` and fetch the three external-data
+files declared by `config.json`.
+
+## Run the CUDA variant with Python
+
+The repository also retains the separately validated CUDA-targeted ONNX Runtime
+GenAI graph.
 
 ```bash
-pip install -U "huggingface_hub[cli]"
-hf download webbrain-one/Ling-3.0-tiny-ONNX \
-  --local-dir Ling-3.0-tiny-ONNX
+pip install "onnxruntime-gpu>=1.28.0" \
+  "onnxruntime-genai-cuda>=0.15.2" \
+  "transformers>=4.57,<5"
 ```
 
-## Run with Python
-
 ```python
 import numpy as np
 import onnxruntime_genai as og
@@ -78,14 +126,10 @@ from transformers import AutoTokenizer
 
 model_dir = "Ling-3.0-tiny-ONNX"
 model = og.Model(model_dir)
-tokenizer = AutoTokenizer.from_pretrained(
-    model_dir,
-    trust_remote_code=True,
-)
+tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
 
-messages = [{"role": "user", "content": "Explain why the sky is blue."}]
 input_ids = tokenizer.apply_chat_template(
-    messages,
+    [{"role": "user", "content": "Explain why the sky is blue."}],
     add_generation_prompt=True,
     tokenize=True,
     return_tensors="np",
@@ -94,66 +138,63 @@ input_ids = tokenizer.apply_chat_template(
 
 params = og.GeneratorParams(model)
 params.set_search_options(
-    max_length=int(input_ids.shape[-1]) + 256,
+    max_length=int(input_ids.shape[-1]) + 128,
     do_sample=False,
 )
 
 generator = og.Generator(model, params)
 generator.append_tokens(np.asarray(input_ids[0], dtype=np.int32))
-
 prompt_length = int(input_ids.shape[-1])
+
 while not generator.is_done():
     generator.generate_next_token()
 
-generated = generator.get_sequence(0)[prompt_length:]
-print(tokenizer.decode(generated, skip_special_tokens=True))
+print(tokenizer.decode(generator.get_sequence(0)[prompt_length:], skip_special_tokens=True))
 ```
 
-Set `enable_thinking=True` to use the base model's reasoning mode. The original
-model card recommends `temperature=1.0`, `top_p=0.95`, and `top_k=20` for
-thinking-mode sampling.
+## Architecture and quantization
 
-## Quantization details
-
-- Dense linear weights: symmetric Q4, block size 32 (`MatMulNBits`)
-- Routed expert weights: symmetric Q4, block size 32 (`QMoE`)
-- Embeddings, activations, and recurrent/KV cache states: FP16
-- MoE router weights and routing math: FP32 to preserve expert selection
-- ONNX graph: 235 Q4 `MatMulNBits`, 23 Q4 `QMoE`, 18 `LinearAttention`,
-  18 `CausalConvWithState`, and 6 `GroupQueryAttention` nodes
+- 18 Kimi Delta Attention layers with recurrent and convolution state
+- 6 Multi-Latent Attention layers with KV cache
+- 23 sparse MoE layers with 128 routed experts, top-8 group-limited routing,
+  expert bias, and one shared expert
+- one dense MLP layer at the start of the decoder
+- symmetric Q4/block-32 dense and routed-expert weights
+- FP16 embeddings, activations, recurrent state, and KV cache
+- FP32 MoE router weights and routing math
 
-Runtime memory is higher than the download size because CUDA kernels, weight
-prepacking, activations, recurrent state, and the KV cache also consume VRAM.
+The ONNX graph contains 235 `MatMulNBits`, 23 `QMoE`, 18 `LinearAttention`,
+18 `CausalConvWithState`, and 6 `GroupQueryAttention` nodes.
 
 ## Validation
 
-The complete artifact was loaded and generated with ONNX Runtime GenAI 0.15.2
-and ONNX Runtime GPU 1.28.0 on an NVIDIA GeForce RTX 5090. A deterministic
-short-prompt smoke test completed successfully. This confirms full graph load,
-tokenizer/chat-template integration, recurrent-state discovery, cache handling,
-and execution of the hybrid Q4 CUDA graph.
-
-File hashes are included in `SHA256SUMS`. On Linux, verify the two graph files
-with:
-
-```bash
-sha256sum -c SHA256SUMS
-```
+- The complete CUDA artifact generated successfully on an NVIDIA GeForce RTX
+  5090 with ONNX Runtime GenAI 0.15.2 and ONNX Runtime GPU 1.28.0.
+- The WebGPU repack was verified tensor-for-tensor against the CUDA graph; all
+  4,835,749,912 external tensor bytes are identical.
+- Transformers.js 4.2.0 loads the config, dispatches `Qwen3_5ForCausalLM`, finds
+  all 18 recurrent/conv caches and 6 attention caches, and applies the original
+  Ling tokenizer/chat template.
+- Browser WebGPU generation validation is required before treating this variant
+  as production-ready.
 
 ## Limitations
 
-- This graph targets the CUDA execution provider and uses ONNX Runtime contrib
-  operators that older runtime versions do not provide.
 - Q4 quantization can change outputs and quality relative to the original BF16
   checkpoint. No benchmark parity claim is made here.
-- The advertised context length is architectural; practical context is limited
-  by available VRAM and workload settings.
-- Only a functional smoke test was performed for this conversion. It is not a
-  substitute for the original model's evaluation suite.
+- Browser support, GPU limits, shader compilation time, and memory behavior vary
+  by operating system, browser version, and GPU driver.
+- The advertised 131,072-token context is architectural; practical browser
+  context is limited by available GPU memory.
+- The WebGPU config uses Transformers.js's existing `qwen3_5_text` hybrid-cache
+  adapter solely as a runtime compatibility layer. The underlying graph and
+  weights remain Ling/Bailing Hybrid, preserved in
+  `config_bailing_original.json`.
 
 ## Attribution and license
 
 The model architecture, checkpoint, tokenizer, and chat template are by
-[inclusionAI](https://huggingface.co/inclusionAI). This conversion retains the
-base model's MIT license. Please cite and credit the original project when using
-or redistributing this artifact.
+[`inclusionAI`](https://huggingface.co/inclusionAI). This conversion retains the
+base model's MIT license. Please cite and credit the
+[original Ling-3.0-tiny release](https://huggingface.co/inclusionAI/Ling-3.0-tiny)
+when using or redistributing this artifact.

SHA256SUMS

@@ -1,2 +1,6 @@
 cdd5029c67a427156a92938f8389482e597324f05ff1b08b364e473b9e75da11  model.onnx
 172b0b42e7cca0df85ada36b05d554a3a5a3fd5ad48492ae55b651c1684d796e  model.onnx.data
+e5061a07964a9eb4ce3744b66781ceecae0fd8a18f75a8de87836a0be1ba75d5  onnx/model_q4f16.onnx
+6b507c4b84ee133d5dedf4ad2fda48119628d8637f0e502adf7182b9d9f72c3d  onnx/model_q4f16.onnx_data
+de1da14397accd5b2470a0877fd4c4f4a703131feb1267e5686ff8091a8030c6  onnx/model_q4f16.onnx_data_1
+edba67f32660ad177bfad6dc7f42bef0e5606cef96bfa51040aad550d2a24eb9  onnx/model_q4f16.onnx_data_2

config.json

@@ -1 +1,126 @@
-{"architectures": ["BailingMoeV3ForCausalLM"], "attention_dropout": 0.0, "auto_map": {"AutoConfig": "configuration_bailing_moe_v3.BailingMoeV3Config", "AutoModel": "modeling_bailing_moe_v3.BailingMoeV3Model", "AutoModelForCausalLM": "modeling_bailing_moe_v3.BailingMoeV3ForCausalLM"}, "embedding_dropout": 0.0, "eos_token_id": 156895, "expert_swiglu_limit_list": null, "first_k_dense_replace": 1, "gated_attention_proj_granularity_type": "head_wise", "group_norm_size": 1, "head_dim": 128, "hidden_act": "silu", "hidden_size": 1536, "initializer_range": 0.02, "intermediate_size": 4608, "kda_lower_bound": -5, "kda_safe_gate": true, "kv_lora_rank": 512, "layer_group_size": 4, "linear_silu": true, "max_position_embeddings": 131072, "max_window_layers": 20, "moe_intermediate_size": 512, "moe_router_enable_expert_bias": true, "moe_shared_expert_intermediate_size": 512, "mtp_loss_scaling_factor": 0, "mtp_use_kda": false, "n_group": 8, "no_kda_lora": true, "norm_topk_prob": true, "num_attention_heads": 16, "num_experts": 128, "num_experts_per_tok": 8, "num_hidden_layers": 24, "num_key_value_heads": 16, "num_kv_heads_for_linear_attn": 0, "num_nextn_predict_layers": 0, "num_shared_experts": 1, "output_dropout": 0.0, "output_router_logits": false, "pad_token_id": 156892, "partial_rotary_factor": 0.5, "q_lora_rank": 256, "qk_head_dim": 192, "qk_nope_head_dim": 128, "qk_rope_head_dim": 64, "rms_norm_eps": 1e-06, "rope_interleave": true, "rope_scaling": null, "rope_theta": 6000000, "rotary_dim": 64, "routed_scaling_factor": 2.5, "router_dtype": "fp32", "scale_router_input": false, "score_function": "sigmoid", "scoring_func": "sigmoid", "seq_aux": true, "share_expert_swiglu_limit_list": null, "short_conv_kernel_size": 4, "tie_word_embeddings": false, "topk_group": 4, "topk_method": "noaux_tc", "transformers_version": "4.45.0", "up_proj_norm": false, "use_bias": false, "use_cache": true, "use_kda_lora": false, "use_mla_nope": false, "use_nGPT": false, "use_qk_norm": true, "use_qkv_bias": false, "v_head_dim": 128, "value_norm": false, "vocab_size": 157184, "model_type": "bailing_hybrid", "torch_dtype": "bfloat16"}
+{
+  "architectures": [
+    "Qwen3_5ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "embedding_dropout": 0.0,
+  "eos_token_id": 156895,
+  "expert_swiglu_limit_list": null,
+  "first_k_dense_replace": 1,
+  "gated_attention_proj_granularity_type": "head_wise",
+  "group_norm_size": 1,
+  "head_dim": 192,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 4608,
+  "kda_lower_bound": -5,
+  "kda_safe_gate": true,
+  "kv_lora_rank": 512,
+  "layer_group_size": 4,
+  "linear_silu": true,
+  "max_position_embeddings": 131072,
+  "max_window_layers": 20,
+  "moe_intermediate_size": 512,
+  "moe_router_enable_expert_bias": true,
+  "moe_shared_expert_intermediate_size": 512,
+  "mtp_loss_scaling_factor": 0,
+  "mtp_use_kda": false,
+  "n_group": 8,
+  "no_kda_lora": true,
+  "norm_topk_prob": true,
+  "num_attention_heads": 16,
+  "num_experts": 128,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 16,
+  "num_kv_heads_for_linear_attn": 0,
+  "num_nextn_predict_layers": 0,
+  "num_shared_experts": 1,
+  "output_dropout": 0.0,
+  "output_router_logits": false,
+  "pad_token_id": 156892,
+  "partial_rotary_factor": 0.3333333333333333,
+  "q_lora_rank": 256,
+  "qk_head_dim": 192,
+  "qk_nope_head_dim": 128,
+  "qk_rope_head_dim": 64,
+  "rms_norm_eps": 1e-06,
+  "rope_interleave": true,
+  "rope_scaling": null,
+  "rope_theta": 6000000,
+  "rotary_dim": 64,
+  "routed_scaling_factor": 2.5,
+  "router_dtype": "fp32",
+  "scale_router_input": false,
+  "score_function": "sigmoid",
+  "scoring_func": "sigmoid",
+  "seq_aux": true,
+  "share_expert_swiglu_limit_list": null,
+  "short_conv_kernel_size": 4,
+  "tie_word_embeddings": false,
+  "topk_group": 4,
+  "topk_method": "noaux_tc",
+  "transformers_version": "4.45.0",
+  "up_proj_norm": false,
+  "use_bias": false,
+  "use_cache": true,
+  "use_kda_lora": false,
+  "use_mla_nope": false,
+  "use_nGPT": false,
+  "use_qk_norm": true,
+  "use_qkv_bias": false,
+  "v_head_dim": 128,
+  "value_norm": false,
+  "vocab_size": 157184,
+  "model_type": "qwen3_5_text",
+  "torch_dtype": "bfloat16",
+  "dtype": "float16",
+  "full_attention_interval": 4,
+  "layer_types": [
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention"
+  ],
+  "linear_conv_kernel_dim": 4,
+  "linear_key_head_dim": 128,
+  "linear_value_head_dim": 128,
+  "linear_num_key_heads": 16,
+  "linear_num_value_heads": 16,
+  "mamba_ssm_dtype": "float16",
+  "rope_parameters": {
+    "mrope_interleaved": true,
+    "partial_rotary_factor": 0.3333333333333333,
+    "rope_theta": 6000000,
+    "rope_type": "default"
+  },
+  "transformers.js_config": {
+    "use_external_data_format": {
+      "model_q4f16.onnx": 3
+    },
+    "kv_cache_dtype": {
+      "q4f16": "float16"
+    }
+  }
+}

config_bailing_original.json

@@ -0,0 +1,83 @@
+{
+  "architectures": [
+    "BailingMoeV3ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_bailing_moe_v3.BailingMoeV3Config",
+    "AutoModel": "modeling_bailing_moe_v3.BailingMoeV3Model",
+    "AutoModelForCausalLM": "modeling_bailing_moe_v3.BailingMoeV3ForCausalLM"
+  },
+  "embedding_dropout": 0.0,
+  "eos_token_id": 156895,
+  "expert_swiglu_limit_list": null,
+  "first_k_dense_replace": 1,
+  "gated_attention_proj_granularity_type": "head_wise",
+  "group_norm_size": 1,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 4608,
+  "kda_lower_bound": -5,
+  "kda_safe_gate": true,
+  "kv_lora_rank": 512,
+  "layer_group_size": 4,
+  "linear_silu": true,
+  "max_position_embeddings": 131072,
+  "max_window_layers": 20,
+  "moe_intermediate_size": 512,
+  "moe_router_enable_expert_bias": true,
+  "moe_shared_expert_intermediate_size": 512,
+  "mtp_loss_scaling_factor": 0,
+  "mtp_use_kda": false,
+  "n_group": 8,
+  "no_kda_lora": true,
+  "norm_topk_prob": true,
+  "num_attention_heads": 16,
+  "num_experts": 128,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 16,
+  "num_kv_heads_for_linear_attn": 0,
+  "num_nextn_predict_layers": 0,
+  "num_shared_experts": 1,
+  "output_dropout": 0.0,
+  "output_router_logits": false,
+  "pad_token_id": 156892,
+  "partial_rotary_factor": 0.5,
+  "q_lora_rank": 256,
+  "qk_head_dim": 192,
+  "qk_nope_head_dim": 128,
+  "qk_rope_head_dim": 64,
+  "rms_norm_eps": 1e-06,
+  "rope_interleave": true,
+  "rope_scaling": null,
+  "rope_theta": 6000000,
+  "rotary_dim": 64,
+  "routed_scaling_factor": 2.5,
+  "router_dtype": "fp32",
+  "scale_router_input": false,
+  "score_function": "sigmoid",
+  "scoring_func": "sigmoid",
+  "seq_aux": true,
+  "share_expert_swiglu_limit_list": null,
+  "short_conv_kernel_size": 4,
+  "tie_word_embeddings": false,
+  "topk_group": 4,
+  "topk_method": "noaux_tc",
+  "transformers_version": "4.45.0",
+  "up_proj_norm": false,
+  "use_bias": false,
+  "use_cache": true,
+  "use_kda_lora": false,
+  "use_mla_nope": false,
+  "use_nGPT": false,
+  "use_qk_norm": true,
+  "use_qkv_bias": false,
+  "v_head_dim": 128,
+  "value_norm": false,
+  "vocab_size": 157184,
+  "model_type": "bailing_hybrid",
+  "torch_dtype": "bfloat16"
+}

onnx/model_q4f16.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5061a07964a9eb4ce3744b66781ceecae0fd8a18f75a8de87836a0be1ba75d5
+size 811364

onnx/model_q4f16.onnx_data

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b507c4b84ee133d5dedf4ad2fda48119628d8637f0e502adf7182b9d9f72c3d
+size 1984862080

onnx/model_q4f16.onnx_data_1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de1da14397accd5b2470a0877fd4c4f4a703131feb1267e5686ff8091a8030c6
+size 1943274240

onnx/model_q4f16.onnx_data_2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edba67f32660ad177bfad6dc7f42bef0e5606cef96bfa51040aad550d2a24eb9
+size 907613952

tokenizer_config.json

@@ -2110,5 +2110,6 @@
   "model_max_length": 131072,
   "pad_token": "<|endoftext|>",
   "tokenizer_class": "PreTrainedTokenizerFast",
-  "trust_remote_code": true
+  "trust_remote_code": true,
+  "chat_template": "{#- Bailing V3 chat template -#}\n{#- Supports: thinking option, tool calling -#}\n\n{#- ==================== thinking option normalization ==================== -#}\n{%- if enable_thinking is defined %}\n{%- if enable_thinking %}\n{%- set thinking_option = 'on' %}\n{%- else %}\n{%- set thinking_option = 'off' %}\n{%- endif %}\n{%- elif thinking_option is not defined %}\n{%- set thinking_option = 'on' %}\n{%- endif %}\n\n{#- ==================== preserved thinking  ==================== -#}\n{% set preserved_thinking = true %}\n\n{#- ==================== system message ==================== -#}\n{{- '<role>SYSTEM</role>' }}\n{%- if tools %}\n    {%- if messages[0].role == 'system' %}\n        {{- messages[0].content + '\\n' }}\n    {%- endif %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nIf none of the functions can be used, point it out. If the given question lacks the parameters required by the function, also point it out.\\nIf you need to use a function, for each function call, output the function name and arguments within the following XML format:\\n<tool_call>{function-name}\\n<arg_key>{arg-key-1}</arg_key>\\n<arg_value>{arg-value-1}</arg_value>\\n<arg_key>{arg-key-2}</arg_key>\\n<arg_value>{arg-value-2}</arg_value>\\n...\\n</tool_call>\\n\" }}\n    {%- if messages[0].role == 'system' and messages[0].content is string and ('detailed thinking on' in messages[0].content or 'detailed thinking off' in messages[0].content) %}\n        {{- '<|role_end|>' }}\n    {%- else %}\n        {{- 'detailed thinking ' + thinking_option + '<|role_end|>' }}\n    {%- endif %}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {%- if 'detailed thinking on' in messages[0].content or 'detailed thinking off' in messages[0].content %}\n            {{- messages[0].content + '<|role_end|>' }}\n        {%- else %}\n            {{- messages[0].content + '\\n' }}\n            {{- 'detailed thinking ' + thinking_option + '<|role_end|>' }}\n        {%- endif %}\n    {% else %}\n        {{- 'detailed thinking ' + thinking_option + '<|role_end|>' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if message.content is string %}\n        {%- set content = message.content %}\n    {%- else %}\n        {%- set content = '' %}\n    {%- endif %}\n    {%- if message.role == \"user\" %}\n        {{- '<role>HUMAN</role>' + message.content + '<|role_end|>' }}\n    {%- elif message.role == \"system\" and not loop.first %}\n        {{- '<role>SYSTEM</role>' + message.content + '<|role_end|>' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is string and message.reasoning_content != '' %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in content %}\n                {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n                {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- if preserved_thinking or loop.index0 > ns.last_query_index %}\n            {%- if reasoning_content != '' %}\n                {{- '<role>ASSISTANT</role>' + '\\n<think>' + reasoning_content.strip('\\n') + '</think>' + content.lstrip('\\n') }}\n            {%- else %}\n                {{- '<role>ASSISTANT</role>\\n<think></think>' + content }}\n            {%- endif %}\n        {%- else %}\n            {{- '<role>ASSISTANT</role>\\n<think></think>' + content }}\n        {%- endif %}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- set tc = tool_call %}\n                {%- if tool_call.function %}\n                    {%- set tc = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>' + tc.name }}\n                {% set _args = tc.arguments %}\n                {%- for k, v in _args.items() %}\n                    {{- '<arg_key>' + k + '</arg_key>' }}\n                    {{- '\\n<arg_value>' }}\n                    {%- if v is string %}\n                        {{- v }}\n                    {%- else %}\n                        {{- v | tojson(ensure_ascii=False) }}\n                    {%- endif %}\n                    {{- '</arg_value>' }}\n                {%- endfor %}\n                {{- '\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|role_end|>' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<role>OBSERVATION</role>' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|role_end|>' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n\n{#- ==================== generation prompt ==================== -#}\n{%- if add_generation_prompt %}\n    {{- '<role>ASSISTANT</role>' }}\n    {%- if thinking_option == 'on' %}\n        {{- '\\n<think>' }}\n    {%- elif thinking_option == 'off' %}\n        {{- '\\n<think></think>' }}\n    {%- endif %}\n{%- endif %}"
 }